ML_MODEL.md

ML Models Documentation

This document describes the Machine Learning models used in the chat application for:

• Spam Detection
• Toxic Message Detection

---

1. Spam Detection Model

Dataset

• File: spam.csv
• Columns used:
  • v1 → label (ham/spam)
  • v2 → message text

Data Preprocessing

• Renamed columns to label and message
• Converted labels:
  • ham → 0
  • spam → 1
• Removed unnecessary columns

Feature Extraction

• Technique: TF-IDF Vectorization
• Parameters:
  • stop_words = 'english'
  • max_features = 5000

Model Used

• Algorithm: Multinomial Naive Bayes

Training

• Train/Test Split:
  • 80% training
  • 20% testing
• Random State: 42

Evaluation Metrics

• Accuracy Score
• Classification Report (Precision, Recall, F1-score)

Prediction Function

def predict_spam(text):
    text_vec = vectorizer.transform([text])
    prob = model.predict_proba(text_vec)[0][1]
    return prob

Example Predictions

• "Free money click now!!!" → High spam probability
• "Let's meet tomorrow" → Low spam probability

Model Export

• spam_model.pkl
• spam_vectorizer.pkl

2. Toxic Message Detection Model

Dataset

• File: data.csv
• Columns used:
  • comment_text
  • target

Data Preprocessing

• Created binary label:
  • toxic = 1 if target > 0.5
  • else 0
• Selected relevant columns:
  • comment_text, toxic

Feature Extraction

• Technique: TF-IDF Vectorization
• Parameters:
  • stop_words = 'english'
  • max_features = 10000

Model Used

• Algorithm: Logistic Regression
• Parameters:
  • max_iter = 1000

Training

• Train/Test Split:
  • 80% training
  • 20% testing
• Random State: 42

Evaluation Metrics

• Accuracy Score
• Classification Report

Prediction Function

def predict_toxicity(text):
    text_vec = tfidf.transform([text])
    prob = model.predict_proba(text_vec)[0][1]
    return prob

Example Predictions

• "You are stupid" → High toxicity
• "Have a great day" → Low toxicity

Model Export

• toxic_model.pkl
• toxic_vectorizer.pkl

Integration in Chat App

Both models are used for real-time message analysis.

Workflow

1. User sends message
2. Message is vectorized
3. Passed to ML model
4. Probability score generated
5. If threshold exceeded:
   • Flag as spam/toxic
   • Take action (warn, block, or filter)

Future Improvements

• Use deep learning models (LSTM or Transformers)
• Improve dataset quality
• Add multilingual support
• Real-time model retraining
• Context-aware toxicity detection

Files

• spam_model.pkl
• spam_vectorizer.pkl
• toxic_model.pkl
• tfidf_vectorizer.pkl

Feature	Model	Technique
Spam Detection	Naive Bayes	TF-IDF
Toxic Detection	Logistic Regression	TF-IDF